Visual On-line Learning in {Multi-Camera} Networks

Abstract

The detection of persons and other objects is an important task inautomatic visual surveillance systems. In recent years, both advancedrepresentations and machine learning methods have proven to yield high recognition performancewhile keeping false detection rates low. However, these approaches require ahigh amount of labeled training data, which is not available for most practicalapplications. Thus, the required labeled samples are usually obtained by hand-labeling, which is a tedious and time-consuming task. Moreover,building general detectors, that are applicable to a wide range of scenes, lead tohigh model complexity and, thus, unnecessary complex detectors. In thischapter, we motivate online learning (i.e., boosting for feature selection) in orderto train fast, scene-dependent, adaptive object detectors. In order to reducethe human effort and to increase the robustness we propose a multi-cameraframework, where each camera holds a separate classifier, which is specified for itsscene or view point. To train and improve these classifiers we incorporateknowledge from other cameras in a co-training manner. The power of our approach is demonstrated in various experiments on publicly available data as well ason data generated in our lab.